The Neocloud Thesis Breaks as Hyperscalers Weaponize Overcapacity

Meta's entry into the GPU cloud market signals that hyperscalers have enough excess AI compute to undercut specialized players; the scarcity narrative that justified neocloucds has collapsed.

CoreWeave's $20–25 billion in capital raises two weeks ago seemed to validate the neocloud thesis: the world needs more AI compute, and specialized providers would capture the gap between hyperscaler capacity constraints and AI lab demand. The capital markets believed it. Then Meta Compute launched, and CoreWeave crashed 25% in fourteen days—including an 11% drop on Meta's announcement. Nebius fell 13%, IREN 6.5%. The neocloud trade unraveled in real time.

The mechanism is straightforward and brutal. Hyperscalers (Meta, Google, Microsoft, Amazon) have committed $80 billion-plus in AI capex from single backers and are staring at utilization problems: their own workloads—search, ads, recommendations—cannot absorb all the GPU capacity they're building. Rather than write down stranded assets, they're leasing excess GPUs to AI labs and startups at scale. Meta's deal with Anthropic was reportedly worth $10 billion. When the largest cloud platforms can field retail GPU offerings, neocloucds—which won on the premise of filling a supply gap—become margin compression under hyperscaler pricing power and customer lock-in.

The supply-side math amplifies this collapse. SK Hynix's $26.5 billion equity raise and CEO warning that 2027–2030 will be "the worst year for memory shortage" signal that HBM, not capacity, is the binding constraint. Hyperscalers control Nvidia allocation, TSMC priority, and chipmaker long-term contracts. Neocloucds, by definition, compete for commodity capacity *after* tier-1 players cherry-pick. ASML's repeated guidance raises and TSMC's $265 billion US commitment show fab capacity is scaling—but the real constraint is hyperscaler overcapacity in the GPU cloud market crushing neocloud margins.

The financing structure compounds risk. CoreWeave's $24.9 billion in debt features circular arrangements with Nvidia customers—a structure that works only if AI capex stays exponential. BIS and institutional investors are now flagging AI infrastructure as "debt-driven excess," a replay of past tech bubbles. If demand cools or customer lock-in breaks, the debt stack at neocloucds becomes a solvency crisis for credit-dependent leverage and pension funds.

Watch: whether this is temporary (hyperscalers leasing transient overcapacity) or structural (hyperscalers permanently owning the GPU cloud market). If the latter, neocloucds face consolidation, single-digit margins, or defensible niches (latency-sensitive inference, geographic arbitrage). The next six weeks—Q2 earnings season—will show if hyperscaler utilization improves or the overcapacity thesis holds and continues to price neocloucds out.